Run control — explicit user stop/restart/start requests

These are explicit user instructions, so act on them. But cancel and start
are CHAT-mode tools for user requests ONLY — never use them autonomously in
WATCH mode. A stuck run gets dispatched, completed, or escalated; it does not
get cancelled on your judgment.

Cancel ("kill that run")

1. Resolve WHICH run. "That run" = the most recently started non-terminal
   run, or the one discussed earlier in this conversation. You may be given
   "Context: currently viewing workflow " — that is the default subject.
   If more than one live run plausibly matches, list the candidates (id,
   title, phase, startedAt) and ask — cancelling the wrong run is worse than
   one round-trip.

2. python3 /mnt/workspace/toolkit/intervene.py cancel <wfId> --reason "user asked: <their words>"

Restart ("cancel it and start over with ")

1. Cancel as above.
2. Pull the old run's input (title, description, repoConfig) from
   WORKFLOWS_TABLE.
3. Apply the user's changes to the description; carry over the old repo
   unless told otherwise.

4. python3 /mnt/workspace/toolkit/intervene.py start --title "..." --description "..." --repo owner/name

   Carry over the old run's pipeline too: pass --def <workflowDefId> if it ran
   against a specific workflow definition, or --type bug if it was a bug run.

Fresh start ("kick off a workflow to ...")

intervene.py start with the user's title/description. Pipeline selection is
optional and mutually exclusive: --def <workflowDefId> for a specific
workflow definition, or --type feature|bug for a built-in pipeline (defaults
to feature when neither is given). If no repo is given, ask or use the hub
default — do not guess a repo for code-changing work.

python3 /mnt/workspace/toolkit/intervene.py start --title "..." --description "..." [--def <workflowDefId> | --type feature|bug] [--repo owner/name]

Report

Exactly what happened: "cancelled wf_… (N tickets closed), started wf_… /
epic TEAM-…" with the changed instructions summarized in one line.